(CW:AI) Discussion of LLMs and other kinds of text generators ("AI"), especially of toki pona

[quiet geyser]

well kijetesantakalu is a type of animal. thus force of nature

also i feel like kijetesantakalu is a very specific combination of letters that if there's anything else named kijetesantakalu, it's named after the nimi

[naive light]

toki a, kulupu lili o
mi weka e ken pana pi sitelen mun ⭐ tan tomo ni
ni li tan ala pakala sina li tan ni: tomo ni la sina ken len ala e sitelen ilo. taso tomo #pona-kulupu la sina o len e ona.
ni li taso. o musi pona

[quiet geyser]

rip this post 😔

also what definition of expert are they using

[rare crescent]

Better than randomly typing bullshit

[heavy shard]

It feels like it’s mocking me

[unkempt maple]

is calling everything a kijetesantakalu or a palisa better than randomly typing bullshit? /lh

[snow echo]

yes

[glass iris]

sona a. ni li pona tawa mi. mi lukin e toki pi tomo ni lon tomo #pona-kulupu la, mi pilin sama sina. toki li tan ilo pi sona ilo la, ona o len. taso ona li ken ala len lon tomo ante ni.

[twilit stag]

WBR seems to always get the words mixed up, it's implying that toki pona created Sonja lang in the second sentence

toki pona li mama pi jan Sonja

[silver ember]

real

[opal hedge]

seme li wile toki insa?

o mu insa o mu kalama kin

[dense nexus]

trained a toki pona classifier, combining that with Kekan San's toki pona tokenizer, to make a really reliable toki pona filter, retaining only toki pona posts from Bluesky in real time.

the toki pona classifier model is available here, (pretty tiny at 118M parameters: https://huggingface.co/johnpaulbin/toki-pona-classifier-v2)

[heady otter]

there is already sona toki (also by the mr. regnal ordinal the third danielson affectionately "gregdan" himself) which does this without AI
i'd imagine it to be faster then this approach

[dense nexus]

wheres the link to sona toki?

[polar orbit]

https://github.com/gregdan3/sona-toki

[naive light]

wait, did you not know about my library? we were talking about it in DMs a couple days ago tho

[dense nexus]

oops 😭 i tottaly forgotten the name of your repo

[heady otter]

yeah when i saw the tp filter on bluesky, i thought it used sona toki

[dense nexus]

it uses sona toki plus my project / model

[sonic leaf]

i wanna talk, i'll talk about sticks.
i wanna wait, i wanna request.
i wanna wait for ears.
i wanna eat a pear.
i wanna fuel my wants.
i wanna talk about talking.
i wanna move my wants.
i wanna look into the sun.
i wanna wait for myself.
i wanna talk about my wants.

[meager jackal]

this is actually a fire jan moose-awi verse

[tropic nova]

ILO IKE toki li Inli ala

[ruby eagle]

Is there a collection of ai training data for pronouncing tp

[hollow osprey]

mi awen e toki [pona] mi lon tenpo weka https://drive.google.com/file/d/14bNf2Ler9cOIGbAsrfceAbU8W29TpFkQ/view

ante la · jan [2swap] li pana e toki kalama mute lon ma [Tatoeba] · taso sina wile pana e ona tawa ilo la o toki tawa ona

[ruby eagle]

Thx! Also, Im not fluent in tp yet, could you translate that pls?

Nvrmnd i figured out what you said

[full narwhal]

thanks

[severe plume]

there's the Mozilla Common Voice collection

[lucid niche]

for the record, he's not a moose. there's a video on this

mijun Niwe (ping reply) li penpo ↩️

[Reply to:](#1162756469771472988 message) this is actually a fire jan moose-awi verse

there's https://commons.wikimedia.org/wiki/Category:Lingua_Libre_pronunciation-tok although I'm not sure if that can be used in AI training or not

jan Pa ↩️

[Reply to:](#1162756469771472988 message) Is there a collection of ai training data for pronouncing tp

(if you find out, let me know! some of my voice is in there, and I'm curious if it can)

[hollow osprey]

[ilo penpo o lukin ala]
on the one hoof public domain means no copyright whatsoever, and a lot of research tts has been trained on public-domain audiobooks
on the other hoof it might come up against personality rights? e.g. a few contributors to mozilla common voice (also cc0) objected to having their data used for tts
on the third hoof the official opinion of creative commons is that training on copyrighted data is fair use
and if i may lie down, on the fourth hoof will single words even be sufficient for whatever project jan Pa has in mind?

[meager jackal]

This video was funny but I saw a comment by Kurz23 which saddened me because how can someone be open minded enough to lean toki pona but not to accept tonsi

[meager jackal]

on the … hoof
poni li lon

[midnight cape]

Error code: a733a13d28a94234a863713ac11d5eb8

Internal error occurred

For support, please send the error code above in #bug-reports-and-errors on the support server (click to join) with a description of what you were doing at the time.

a733a13d28a94234a863713ac11d5eb8 · Dec 3, 2024, 11:42 PM

[dense nexus]

whats your main goal?

[meager jackal]

kin la i love "and if i may lie down" for the 4th hoof

[cursive ore]

chat gpt just seems to make stuff up when you ask it for stuff about toki pona

[opal hedge]

mokon
moku lon

[mossy pagoda]

Sorry to kinda necropost, I had a conversation with jan Sonja if I remember correctly on why we shouldn't have a luck/choice word! I wonder what the AI would do if presented with feedback from the language's creator

[hollow osprey]

mi kin li wile sona e tan

[mossy pagoda]

Found it! Keep in mind this was over 9 months ago, so I'm a far better tokiponist hopefully! Ignore anything dumb I say!

[remote dove]

kon: am i a joke to u

[meager jackal]

this is like the "three gry" riddle.

"there are three words in toki pona which end in -on. the first is lon, and the second is kon. what is the third?

mokon"

[rare crescent]

I like how it says "unofficial" without explaining what "official" is

makes it sound like there's a high register of toki pona spoken by the pu government of ma pona and then there are countless unofficial dialects with a bajillion words, derivations, and adaptations

[severe plume]

iirc Sonja uses "official" to mean "by Sonja"

or like "Sonja's version of toki pona"

i forget what exactly, but it has to do with her specifically

[wet pewter]

i mean for all we know it could be correct, maybe people do use “mokon” but since it isn’t commonly used not many people know /j

[rare crescent]

yeah we all know poetic forms, derived and alternative forms, and dialectal words and spelling is so common within toki pona /j

[glass iris]

I like this idea. Or at least, I like the idea of telling people this is how toki pona works. "Can I say X in toki pona" "it's fine with me, but I just checked in with the grand Vizier of the permanent ruling council of pu, and they said no. Sorry. "

[rare crescent]

"you may switch to a lower-class dialect, but that'll significantly downgrade your ma pona social status and you'll never get a toki pona job if you don't speak nasin pu"

[glass iris]

I don't like it any more than you do, but that's just what society is like. I'm sorry.

[devout wedge]

love how it completely ignores "ale"

yknow... the toki pona word for "all"?

[hollow osprey]

mi lukin la jan [Sonja] li toki ala anu seme e ni → "nimi pi kon 🎲 o lon ala"

[hollow osprey]

ma [Amazon] la lipu ona li lon kulupu "Official Toki Pona"

[silver ember]

lmao

[rare crescent]

i ride my raccoon to toki pona

seems legit

[sonic leaf]

kule and wile as well

[glass iris]

If only we could star things in this channel because that is a classic

[glass iris]

This is why you never use ChatGPT instead of Google, kids

[rare crescent]

I never understood people who use chatgpt for simple questions instead of a search engine. like, why would you do that? can't you just, well, search? it's not like chatgpt has more information, or more reliable information, or anything else. you are just asking a question to an ai model that just guesses what to answer instead of actually searching the internet. why? what's the point? it's not better in quality. it doesn't provide you with more information. it isn't even easier. you can just go to any search engine, type the question, and probably get a highlighted text on the first page telling you the answer, and maybe think for two seconds, but for simple questions, that's all for the basics. no idea why so many people use chatgpt for that... just ask an actual source instead of an autocomplete program...

[glass iris]

I think it has a lot to do with the fact that most people still don't understand how Generative AI works. I have another theory too, but I'm running between things - it's more philosophical, I'll share it later!

[dense nexus]

Speaking of gpt and searching the web

I do have gpt pro and they have a web search feature. Here I tried "teach me toki pona" it didn't help me directly but it did list some resources... Hey I know that person!

[meager jackal]

"according to chatgpt" you couldn't put 1 google search?? 1 wiktionary search?? (although both google translate and baidu get it wrong because the readings of the characters individually changes whne put together)

嗝屁 btw

luckily not everyone has that mindset

[wet pewter]

sometimes i do it because i have a more complicated question that i feel wouldn’t work well with a search engine, like long sentences with complex phrasing

[rare crescent]

I’m talking about simple questions though. Like, just random facts with a two word answer type questions

[jaunty jungle]

but is or is it not geE pIi

[wet pewter]

ah, for that one i have no idea
imo it’s not even easier to use an AI, it’s easier to just use a search engine

[sonic leaf]

i use chatgpt for things that i don’t know much about, where it’s kind of hard to know where to start. so even for something that might have a technically correct two word answer, if i have no experience in that field, i’d probably ask follow up questions

[real narwhal]

in other words you make it harder for yourself to see if it's wrong

[meager jackal]

it's gěpì

[opal hedge]

a kule kalama li ante lili

[meager jackal]

which is why they were so downvoted. if ai was right it probably wouldn't have happened

[glass iris]

ijo suwi o...

Exciting news! I'm part of team applying for a big federal grant (National Endowment for the Humanities) about Generative AI and the Humanities. I'm heading up the part of the team on AI and Creativity. At a planning meeting this morning, there was a feeling that we should do something about machine translation, particularly from languages with small textual corpora - with an eye to language preservation etc. My colleague, who is Abenaki, said that no way in hell were we letting an LLM anywhere her language, as endangered as it is.

So, I finally outed myself as a jan pi toki pona, and told them all about this thread and our experiments. People were very excited. So, if we get this grant, some of it will be devoted to LLM translations to and from toki pona. As part of that, I'll be teaching the language to a bunch of undergrads, who will then be let loose on a variety of models, and will experiment with prompting etc.

I haven't mentioned this yet to the team - but I'd also like to find a place for people here to participate - since this is probably the most experienced group of people anywhere on this subject. I'll keep you all updated if we get the grant.

[dense nexus]

WOOOHOOO

As for prompting, it could be possible, but despite my best efforts I couldn't get only prompting to work with understanding toki pona.

(tested with Gemini 1.5 pro by scraping very popular toki pona resources)

Seems to take things too literally, as it misses nuances that would only be possible if trained on toki pona corpuses

I'd love to contribute what I know so far

[dense nexus]

What corpuses are available that are in toki pona, but are meant to be for education? (i.e., educational subjects, can also be toki pona subjects)

[opal hedge]

lipu tenpo is educational

[dense nexus]

Ty for the resource

[glass iris]

I think one outcome of the study will be precisely that: that prompting doesn't help very much. We'll also try more systematically to figure out the kinds of mistakes different LLMs tend to make, and see if there is any possible way to figure out the cause. (Some versions of ChatGPT sound as if they're writing their response in English, and then translating literally- but how can that be, given the n-gram token production mechanism of transformer architectures?) And it will be interesting to think why Claude is so much better than any other.

[glass iris]

I thought too that translating lipu tenpo articles to English would be a good test case. The generous license allows that kind of use. (And in any case, I have half a dozen articles in there that I'd be happy to set them loose on).

[lucid niche]

I see, good to know! there are actually a few sentences in there (including the entire lyrics to soko soweli by kqa, broken up by sentence. I am constantly questioning my systemmates' decisions), but yeah for the most part it's words
-# (also I love the four hooves phrasing)

⦅wile sina⦆ Kita [⥀·⫞·] li tpt ↩️

[Reply to:](#1162756469771472988 message) [ilo penpo o lukin ala]
on the one hoof public domain means no copyright whatsoever, and a lot of re…

yeah...jan Kuse is a whole deal. it's alright though, we love jan Usawi and her antlers

mijun Niwe (ping reply) li penpo ↩️

[Reply to:](#1162756469771472988 message) This video was funny but I saw a comment by Kurz23 which saddened me because how can someone be open…

mokon 😭 if you're going to skip kon, at least do like lokon or something

jan Pen ↩️

[(click to see attachment)](#1162756469771472988 message)

Official Toki Pona refers to her book series

waso Keli ↩️

[Reply to:](#1162756469771472988 message) i forget what exactly, but it has to do with her specifically

and if you really are fed up of Google, next time try out your phone's text-replacement algorithm. it's basically the same stuff

soweli Lapate la ale li suwi ↩️

[Reply to:](#1162756469771472988 message) This is why you never use ChatGPT instead of Google, kids

[dense nexus]

personally i think simply that claude has the most propritary dataset ever, it knows things that i would imaging being only capable of knowing via a github scrape

probably some very clever web crawling + not caring about licenses

[lucid niche]

whoa, very cool!! teach them well

soweli Lapate la ale li suwi ↩️

[Reply to:](#1162756469771472988 message) ijo suwi o...

Exciting news! I'm part of team applying for a big federal grant (National Endowment …

[dense nexus]

however writing a full report would be so so so interesting to read!!! im more than exited for the future researching to be done.

[dense nexus]

does the license allow using it for LLM training? i ask since i would really love to get knowledge transferring with llama 3.1 (8B) and toki pona

would be more than useful for translation specifically

[quiet geyser]

is official her english title

official Sonja Lang

wouldn't this make unu official

[orchid veldt]

well, usually anything by Sonja that isn't the language itself
like official books, official website, official shirt store...

[glass iris]

You should double check the license. But when I asked the jan pali about using it to fine tune a language model, they said sure, as long as you identify publicly the source. The license should specify exactly how you're supposed to do that.

It's a very good source, because everything is written by proficient tokiponists, and then proofread by other proficient members of the community. It's as close as you'll get to guaranteed "correct" language.

[dense nexus]

thats awesome, ill make sure to double check

ty!

[heavy shard]

sorry to detract from this quite goated conversation but

i just wanna say

i think everyone should at some point try forcing google gemini to engage in a toki pona conversation with them

[quiet geyser]

maybe it's official because she has an office

[glass iris]

la, seme li kama?

[quiet geyser]

ni li tawa sina taso, jan o

[lucid niche]

ah, so nusun is official

jan Ke Tami [K · ][𐌝 · ρ : ] ↩️

[Reply to:](#1162756469771472988 message) well, usually anything by Sonja that isn't the language itself
like official books, official website…

as is tankala

not to mention mulapisu

[quiet geyser]

toma is in ku so it's undeniably official

[lucid niche]

of course

[heavy shard]

here are some things i have experienced trying to have a conversation with google gemini in toki pona:

• at first, it vehemently refused to get the picture. it just replied with a thing it thought my toki pona meant (it was wrong)
• when i eventually told it to speak toki pona with me, it began speaking nearly entirely unintelligible toki pona. i'm talking... like... "misuse of particles as intensifiers" bad. it speaks with each line separated by line breaks and a side-by-side translation.
• somehow, when sending the prompt "sina pana e toki tu tu. toki tu tu la ale li pakala. toki pona sina li ike mute. o kama sona." (this isn't formatted fantastically but you get the picture) it assumed i was talking in Marathi and it responded back entirely in Marathi.
• when telling it in toki pona that it did this, it then began to spit out a barrage of sentences in an endless loop (most if not all wrong), only stopping presumably when it hit its character limit.
• when i tried again to engage in toki pona conversation after this, it continued to just express that it was good and wanted to talk to me. no attempts to talk about anything other than "are you good" or "can you speak toki pona" worked. keep in mind most of their sentences were still pretty wrong.

[quiet geyser]

ok just saw the etymon

tbh should be nujun

especially since the <j> is like /ɟʝ/

[glass iris]

I guess even less toki pona in its training material than GPT

[heavy shard]

ike a

[lucid niche]

😭

give it https://nimi.li/kijetesantakalu or something

[quiet geyser]

i mean all tp words are made up

the lang is a con

lang

[hollow osprey]

[ilo penpo o lukin ala]
it's the same license as wikipedia, and most people seemed fine with e.g. translation models being trained on wikipedia
i think teeechnically if you wanted to super-comply you'd have to train a model purely on licensed data
but also the creative commons organization has stated that they believe neural network training to be fair use

[dense nexus]

well shoot

makes sense tho

i dont know if i can create a model from scratch using -only- licensed data

[hollow osprey]

[ilo penpo o lukin ala]
the only such language model i know of is santacoder, which notably still obeyed opt-out requests despite mit being non-revocable
which btw i expect more people to object based on "i don't want this to be used to train ai" than "you didn't follow the share-alike clause fully"

[dense nexus]

santa coder is good but also really big, also unsure how well it is knowladge wise, ill have to do some digging

[hollow osprey]

The SantaCoder models are a series of 1.1B parameter models
ilo toki pi tenpo lon la ni li ala :p

[dense nexus]

oops the model size could just be the optimizers not being removed 😭

on disk it shows 4.6 gb but thats usually meant for larger models, probably something with the checkpoint

[hollow osprey]

? 1.1b * 32bit ≈ 4.6gb anu seme

[dense nexus]

oh is it stored in 32bit? oops

[hollow osprey]

ken

[dense nexus]

i might have to finetune santacoder as you suggested, as my searching shows no other models that have been trained on specifically permissible data

[hollow osprey]

sina kute ala e mi tan seme 😭

i expect more people to object based on "i don't want this to be used to train ai" than "you didn't follow the share-alike clause fully"

[dense nexus]

oh sorry im feeling really tired

okay ill most likely settle on llama 3.1 8B, seemed to have the most impressive adaptation to toki pona by far. things i want to try are:

1. including synthetic Q&A inside the dataset directly after data entries
   • iirc, having finetuning data within the pretraining data helps the model adapt better
   • similar to RAG, this should help the model connect the idea of *translation* more efficiently
2. include lipu tenpo in some way for a general knowladge improvement
   • current toki pona llama 3B model finetune seems to understand basic intentions of toki pona words, but fails to connect conecpts that were only displayed in english, failing to utilize the concepts in toki pona conversations
   • im thinking that this method could heavily help with learning transfering
   • currently, the 2nd biggest corpus (only behind hecko's chat logs) is a toki pona nltk corpus, which seems to only contain more mundane (i,e, less educational) concepts (as seen in emails / everyday conversations)
   • i really want the model to interpret english and toki pona better
3. providing the model high quality toki pona learning resources
   • even claude hallucinates with toki pona knowlage, despite translating it fluently, it fails with facts of the language, as if it were trained on toki pona texts but not toki pona tutorials / guides / info

[hollow osprey]

currently, the 2nd biggest corpus (only behind hecko's chat logs)
seme a‽ mi wawa ni‽

poka la · sina moku ala moku tan https://wikisource.org/wiki/Main_Page/tok

[dense nexus]

I havent used the toki pona wikipedia if that's what you're reffering to, and yes, half of my pretraining dataset was your chat logs lol

i really really really wonder how claude created their dataset for sonnet 3.5

[hollow osprey]

a mi pakala · https://wikisource.org/wiki/Main_Page/toki_pona

lipu pi jan mute li lon ona li lon nasin pana CC

[dense nexus]

Oooh

thank you for pointing that out

[dense nexus]

which website would you recommend i use to provide the model the best insight into learning toki pona and its concepts?

[hollow osprey]

https://sona.pona.la/ li ken

[dense nexus]

thank you!!

[hollow osprey]

a a ilo [Codeium] li sona e ni → toki [pona] li lon · (mi pana e nimi toki mute pi toki [pona] e nimi "lipu")

[orchid veldt]

anything that isn't the language itself

And Sonja can make something and not include it in the "official"

[heady otter]

me when I spend my taxes on Toki Pona /nm /positive

[lucid niche]

wh-

a, pakala mi. ale la mi musi taso

jan Ke Tami [K · ][𐌝 · ρ : ] ↩️

[Reply to:](#1162756469771472988 message) anything that isn't the language itself

I thumbs-downed it and it changed to this

[glass iris]

Every penny spent on toki pona is one less penny going to the military industrial complex!

[heady otter]

agreed

[glass iris]

ilo li ike a. Just so bad. I despair a little.

[lucid niche]

ilo li ike lon

mi o weka e ona ni

...pakala. lukin la mi kama ken ala

[glass iris]

It's interesting that it's not stochastic at all. We can get exactly the same response

ChatGPT (4o):

Marginally better - at least somewhat coherent. But it should of course have said ma pi nena pimeja (LLMs rarely get "pi" placement correct; and don't even seem to know about rarer words like "nena")

[wet pewter]

i feel like pi isn’t strictly necessary there

it works with or without it

only difference is “place of black mountain(s)” or “black mountainy place” i guess

[tawdry bone]

sona mi la nimi official li toki e esun lipu ona

[glass iris]

Yes, those are slightly different - but both could be meaningful in this context.

ilo [Claude] kin li ike:

(this is the free haiku model)

I love how it bullshits an explanation at the end about what is "more natural" in toki pona.

[silver ember]

no way

"mustuy" is just straight from finnish

[glass iris]

Wow - that's interesting!

[silver ember]

"musta" is black in finnish

[glass iris]

And pimeja is also from Finnish - what an interesting mistake

[silver ember]

thinking toki pona is finnish is a mistake that happens to actual humans too so it's not that surprising i guess

[glass iris]

But it is odd to substitute a toki pona word that is actually from Finnish, with another made-up word that is also based on a Finnish word

[orchid veldt]

esun len kin li ni ala ni

[silver ember]

that's hilarious

[glass iris]

facepalm in a kind of burn them all to the fucking ground mood today.

[silver ember]

wait wrong reply

[lucid niche]

I'm pretty sure it generates a response the first time something is searched

soweli Lapate la ale li suwi ↩️

[Reply to:](#1162756469771472988 message) It's interesting that it's not stochastic at all. We can get exactly the same response 📎

[silver ember]

"pimeja is derived from the english word "primitive"" is extremely funny tho

[lucid niche]

and then serves that same one when it's searched again

[glass iris]

Agreed - prime LLM bullshit

[glass iris]

Oh ni li sona musi - I didn't know that

[lucid niche]

(except if that one's given a thumbs down, in which case it seems to generate a new one? but that caps at one regeneration)

quite a lot better, actually!

soweli Lapate la ale li suwi ↩️

[Reply to:](#1162756469771472988 message) ChatGPT (4o): 📎

but Montenegro's not the native name

for me, without pi that implies that the place is black

kasi Majeli ↩️

[Reply to:](#1162756469771472988 message) it works with or without it

[glass iris]

Not sure how to tokiponize Crna Gora

[lucid niche]

Sinakola is pu I think, I've also seen Sunakola

soweli Lapate la ale li suwi ↩️

[Reply to:](#1162756469771472988 message) Not sure how to tokiponize Crna Gora

[dim echo]

Excel

database

...

[glass iris]

Sena Kola?

[lucid niche]

probably any SVnakola is fine

https://en.wiktionary.org/wiki/File:CrnaGora.ogg

I've just suddenly remembered "Thanks, Robert!"

soweli Lapate la ale li suwi ↩️

[Reply to:](#1162756469771472988 message) Agreed - prime LLM bullshit

[glass iris]

I know a spreadsheet isn't exactly a database. But that was the accessible example given in the computer scientist's talk. He meant a static, dependable table of data from which one can reliably extract the same information every time.

[lucid niche]

went and found it

ijo ↩️

[Reply to:](#1162756469771472988 message) i love how it gives your name every time:
"Thanks, Robert!" "Another great question, Robert!" "Preci…

[glass iris]

That was back when I gave a presentation and none of my colleagues knew about toki pona - even though I really wanted to talk about it in that paper! It feels nice to be out of the linguistic closet

[lucid niche]

the closet is a funny metaphor for this

[opal hedge]

a a a mi pilin sama

[lucid niche]

"Mom, Dad...I'm a toki pona speaker. I hope you're okay with it."

[glass iris]

No disrespect meant of course to those who are actually having to struggle with being in the closet

[lucid niche]

a a it's funny to me as a queer person

[dense nexus]

training the new toki pona model:

• base model: llama3.1 8B
• dataset:
  • toki pona bible project
  • hecko's chat messages (only toki pona)
  • lipu tenpo
  • wikisource / toki pona
  • sona pona
  • toki pona nltk corpus
  • synthetic claude 3.5 conversations
  • feedback from the toki pona 3B model

training on a A100, should take less than an hour as it says on the trainer

[tawdry bone]

n ken la esun ale ona li ni

[hollow osprey]

ona li pali wan e ona li pana e ona sama lon tenpo ale anu seme

[dim echo]

mi wile lukin kepeken e ni. taso ilo sona mi li wawa ala

[glass iris]

lukin la ni li nasin ona.

[dense nexus]

don't worry, ill be hosting interactive demos where you can try using the model / provide feedback to it to improve it directly

whoops running out of huggingface storage

[dim echo]

btw, is this dataset public as an actual dataset instead of this list of data sources?

[dense nexus]

ye i can upload the json here if needed

or the raw txt file

@hollow osprey would you be confortable with sharing the dataset with your logs in it?

[hollow osprey]

✅

[dense nexus]

this includes all the sources i listed

jsonl if needed

(jsonl is tokenized for the llama 3 tokenizer, with the context size of 4096)

[dim echo]

imo it should be on hugging face in case anyone will do something related to toki pona

[hollow osprey]

https://wikipesija.org/ li suli li ken pona · taso lipu mute li ike la nnn

[dense nexus]

i will its just i have to work out the description

[dense nexus]

oh gotcha

https://huggingface.co/johnpaulbin/llama3.1-8b-tokipona-toki-1

model available here

example testing it:

[{'role': 'system', 'content': 'You are a helpful assistant.'},
 {'role': 'user', 'content': 'hello!'},
 {'role': 'assistant', 'content': 'Hello! How can I assist you today?'},
 {'role': 'user', 'content': 'how can i say hello in toki pona?'},
 {'role': 'assistant',
  'content': 'In toki pona, you can say "suno o, toki!" or "suno o, toki!" (literally "Hello, speak!" or "Greetings!"). However, a more common way to say "hello" in toki pona is "suno o!" (literally "Sun!"), "jan pona o!" (literally "Friend person!"), or "toki!" (literally "Speak!").'}]

(note i havent done any post finetuning in regards training it with toki pona specifically, im inferencing the "base" model of the toki pona model)

[devout wedge]

not really? a good amount of words are taken from irl languages

iirc kijetesantakalu is taken from finnish

[dense nexus]

translation example

(still the base model)

{'role': 'user',
  'content': '"ma sewi is a server for toki ponists to discuss religion, faith, and spirituality with one another in any language (especially toki pona). It is a place for interfaith discussions, which means that instead of arguing over truth, folk accept that truths can differ for people of different faiths. "'},
 {'role': 'assistant',
  'content': 'Translation: "ma sewi li ilo pi kulupu jan pi toki pona tawa ni: ona li toki e ijo pi nasin sewi, pilin sewi, en pilin pona tawa jan ante. ona li ma toki pi kulupu ante pi nasin sewi. tan ni la jan li toki ala e ni: nasin sewi pi jan wan li lon. jan li toki e ni: nasin sewi li ante tawa jan ante."'}]

another one

  'content': '"Garry\'s Mod, commonly clipped as GMod, is a 2006 sandbox game developed by Facepunch Studios and published by Valve. The base game mode of Garry\'s Mod has no set objectives and provides the player with a world in which to freely manipulate objects. Other game modes, notably Trouble in Terrorist Town and Prop Hunt, are created by other developers as mods and are installed separately, by means such as the Steam Workshop. Garry\'s Mod was created by Garry Newman as a mod for Valve\'s Source game engine and released in December 2004, before being expanded into a standalone release that was published by Valve in November 2006. Ports of the original Windows version for Mac OS X and Linux followed in September 2010 and June 2013, respectively. As of September 2021, Garry\'s Mod has sold more than 20 million copies. A successor, Sandbox, has been in development since 2015."'},
 {'role': 'assistant',
  'content': 'Translation: "jan Gari li pali e ilo musi pi pali mute lon tenpo sike 2006. nimi pi ilo musi ni li \'Garry\'s Mod\' anu \'GMod\'. ilo musi ni li jo e nasin musi wan. taso jan ante li pali e nasin musi ante mute. jan mute li pali e nasin musi \'Trrouble in Terrorist Town\' en \'Prop Hunt\' kepeken ilo musi ni. jan Gari li pali e ilo musi ni kepeken ilo musi pi jan Palu. ilo musi ni li kama lon tenpo sike 2004. taso jan Gari li pali mute e ilo musi ni. tenpo sike 2006 la ilo musi ni li kama lon. ilo musi ni li kama tawa ilo Mac OS X lon tenpo sike 2010. ilo musi ni li kama tawa ilo Linux lon tenpo sike 2013. tenpo sike 2021 la ilo musi ni li kama jo e mani mute. ilo musi sin pi pali mute li kama lon tenpo sike 2015."'}]```

hey all, opening up the toki pona translator here: ~~https://8a2c06126446b95e6f.gradio.live/~~

it should be able to automatically detect which language you are translating from and to

(e.g. entering toki pona will return english and vise versa)

feedback isnt being collected on the demo (but you can still send feedback here) as its more of a test run but feel free to try

[glass iris]

It's really very impressive! But I just hit my first error - do you want it reported somehow? Here?

[dense nexus]

yep here should be okay

and thank you!

[glass iris]

OK (slightly personal) one of my prayers every night is for my friend Hannah, who is a nun, and looks after the cows in her convent - so I always say this "sewi o awen e jan Ana e soweli ona e meli sewi ona" - which the translator gives as "God preserve Anne, her cat, and her husband" - I don't know where it gets "husband" from meli sewi - I mean her "holy sisters" "nuns"

[dense nexus]

oh interesting

i assumed that incorperating the toki pona bible project would cover the use case of that translation, perhaps more context would help? unsure but otherwise ill be sure to work out including more context

maybe have previous prayers within the translation prompt, may help guide the model

[glass iris]

The rest is good - but mije ona or mije olin ona would be husband. meli sewi is clearly "holy woman/women"

[dense nexus]

ah okay

[glass iris]

maybe because meli/mije have become less common in toki pona? idk

[dense nexus]

lol it gave me a run down

[opal hedge]

gender-neutral?

[dense nexus]

its surpisingly good (at least from what i can tell) with handling context

e.g. you can give it discord chat logs that are in toki pona

variable responses though

[rocky stirrup]

interesting

[dense nexus]

lmao

thats in my prompt

[rocky stirrup]

it also seems pretty certain that soweli means cat

[dense nexus]

i think it interprets that by default if you dont include more context in the prompt, but i could be wrong. ill most likely be finetuning it

[dim echo]

nja...

[dense nexus]

ooh i could make the translator provide multiple possible translations + a breakdown

similar to claude

ill try to make a synthetic dataset for that with claude

[dim echo]

mi pilin e ni: ona li sona ala e nimi nja

[tawdry bone]

a ilo li lon li ken pali e ante toki pi pona lili la mi pilin ike tan ken ni: jan li tawa lipu wikipesija li pana e jaki mute tan lipu pi toki inli kepeken ilo ni...

[dense nexus]

that's valid critism, i dont want people to rely on this for use cases like that yet

its not done yet either

[dense nexus]

what specific examples do you have?

[dense nexus]

maybe even preferences like formality, accuracy etc.

[dense nexus]

i think too it would be interesting to pair it with grounding from a search engine

closing the demo for expiramenting with new prompting methods

will reopen soon

[glass iris]

Crazy because the actual word for cat is kisa suwi puwa ko nja

[dim echo]

it's down 😦

[dense nexus]

sorry about that, currently tweeking the backend to see if the model performs better

[dense nexus]

here: https://345694ca9cfb22a2b7.gradio.live/

the model is being prompted differently to produce a breakdown

so results may be slower, however please let me know if translations are more accurate

[dim echo]

what if you include the whole dictionary including less common words in the system prompt? or use some kind of RAG

[dense nexus]

ooh yeah that sounds like a good idea

i could probably do that

[lucid niche]

are Tatoeba / Wikipesija usable?

mega b ↩️

[Reply to:](#1162756469771472988 message) training the new toki pona model:

• base model: llama3.1 8B
• dataset:
  • toki pona bible project
    … 📎

the problem with that is that past the ~140 most commonly used words most words aren't used or understood

mega b ↩️

[Reply to:](#1162756469771472988 message) ooh yeah that sounds like a good idea

so if you tell it that kan, for example, is a word, then it might use kan. this is a problem because barely anyone has used kan since 2002, and few understand it

this is actually a surprisingly good explanation! a man can be "meli" as welll—it's a descriptor that implies femininity. I think @signal obsidian has thoughts on this?

mega b ↩️

[(click to see attachment)](#1162756469771472988 message)

[signal obsidian]

I do! I've been meaning to post that essay at some point lol. But yeah, my contention is that there are no words in Toki Pona for gender. "meli" happens to usually mean "woman" but any gender can be meli.

[lucid niche]

ni li nasin pi pona a

[tawdry bone]

mi pilin e ni: ken la, tenpo ala la ona li pona tawa ni

[dense nexus]

Oh yes! Tatoeba is actually in the dataset already, i forgot to list

[dense nexus]

i might have to agree: i think that even in the mainstream wikipedia.org, using google translate for example to translate english articles into spanish is not as good as actual spanish speakers

[dense nexus]

thats awesome! surprisingly, im pretty sure it learned the concept from one of the wikipedia sources

which is a really amazing occurance, since it means the model "knows" the concept

[signal obsidian]

One example I give in the essay I drafted but never finished is: Suppose two lesbians, both strictly into women, are at a bar. One says "mi alasa e olin sin". The other nods to a cute fem at the end of the bar and says "ona li pona sijelo". The first shakes her head. "ala. mi wile e mije."

[quiet geyser]

njan

[orchid veldt]

n'jan 🎩
n'meli 🎩
n'meli 🎩
n'tonsi 🎩

[lucid niche]

cool!

mega b ↩️

[Reply to:](#1162756469771472988 message) thats awesome! surprisingly, im pretty sure it learned the concept from one of the wikipedia sources

oh @dense nexus I'd highly recommend you feed it this

if you can

it's a comprehensive CC BY-SA 4.0 grammar of toki pona, and explains in detailed prose how virtually every grammar concept works and is used in the common nasin

[dense nexus]

oooo ill include that!

[dense nexus]

translator demo link closed for now, ty for those who gave feedback!!

[meager jackal]

"gender neutral term for woman"

amazing

[lucid niche]

well it can be gender neutral

o lukin e toki pi wan Tansin. kin la a man can be meli too

[quiet geyser]

i guess it's funny since it defines meli as "woman" and "female person" at first rather than "feminine person" or something

[lucid niche]

yeah

[tranquil crow]

If you have access to Claude 3.5 sonnet, then please send me the English translation of this tokipona text

ni o kama pona. nasin tu wan taso li lon tawa wile ni. ona nanpa wan li ni: mi taso li lawa e ijo ale ona. ona nanpa tu li ni: sina taso li ni. ona nanpa tu wan li ni: lawa pi kipisi wan li tawa sina. ona pi kipisi sama ante li tawa mi.

[glass iris]

That seems to me to make perfect sense contextually - and in that case meli/mije can be separate from gender, and instead refer to (I guess?) gender expression. But I don't think that means that meli/mije are no longer gender words. All words in toki pona have that flexibility, and ability to be pushed beyond their semantic space. But if they don't have a clear enough semantic space in the first place, you can't do clever things with them - such as your example.

This is more #sona-musi stuff. But I think examples like yours are precisely the kind of creative and witty use of language that an AI is never going to be able to do.

If we do get this research project funded, I'd love to put together a list of sentences where words are pushed as hard as this out of their normal spaces - as a test to machine translation

[serene acorn]

has anyone figured out why ai loves palisa so much

[glass iris]

I might say meli pi nasin/kule mije

[dense nexus]

"This should be done well. There are only three ways to achieve this goal. The first one is: I alone control all of it. The second one is: you alone do it. The third one is: control of one part goes to you, and control of the other equal part goes to me."

Note: The translation aims to capture the meaning while making it natural in English. Toki Pona is known for its minimalist nature, so some context might be needed for a more precise interpretation.```

[glass iris]

I posted way back on this. A really interesting paper (investigating GPT-J) showed that when you kept pushing that LLM beyond its training resources, it fell back on sexual (especially phallic) and violent imagery. It was like a glimpse into the messy subconscious of the LLM - or perhaps our subconscious, since it's trained on our language. I wonder if that is also happening here, since palisa appears as a phallic word sometimes. (But it doesn't grab onto unpa, oddly)

But given its very limited training in toki pona, it's very easy to push it beyond its limits - so it gets to that toxic state quickly

Here's my original message: #1162756469771472988 message

[lucid niche]

in the same way that a tomo tawa is tawa or an ijo suwi is suwi, a meli mije is by definition mije

jan Nesi ↩️

[Reply to:](#1162756469771472988 message) I'm guessing since you said that they are both strictly into women that by mije they mean a butch. I…

[quiet geyser]

toki palisa

[hollow osprey]

la poki sina li sina anu seme :p

[meager jackal]

yes, because "li sina" can mean "to be yours"

poki mi li mi

[lucid niche]

yes

⦅wile sina⦆ Kita [⥀·⫞·] li tpt ↩️

[Reply to:](#1162756469771472988 message) la poki sina li sina anu seme :p

[twilit stag]

Oh yeah I saw that one

[twilit stag]

Oh yeah because you sent it lmao

[glass iris]

ni a

[dense nexus]

https://huggingface.co/johnpaulbin/llama3.1-8b-tokipona-toki-2

second run of llama3.1 8B

more epochs + trained on synthetic chat data from claude3.5 sonnet

seems to be pretty good with translation, perhaps better than the last model

opening demo really soon

also planning to train a toki pona vision model with the help of hecko

[dense nexus]

this also comes with the ability to chat, but the demo will be focused on translation only for right now

[dense nexus]

added an optional context box

plus an extra context checker if you wish to translate less formally

[glass iris]

Wow! Looking forward to being able to try this out

[dense nexus]

https://1053696d53304e6c25.gradio.live/ Demo now open here

Added the ability to chat in a new tab

(chatting seems to have better translations than the translator feature, but can have worse results in general. however it is also more flexible)

chatting seems to be hallucinating more for some reason

hmm translation seems to vary way more than claude, however prompting it correctly brings good translations

feel free to test out prompting methods to bring the best out the model

this prompt seems to work decently well with translating inside the "Chat" tab:

EXAMPLES OF TRANSLATIONS:
[
{Example 1:
Input: "ona li kama sona e toki pi ma ante lon tenpo suno ni."
Translation: "They are learning a foreign language these days."
},

{Example 2:
Input: "a a a mi wile moku!"
Translation: "Haha, I'm hungry!"
},

{Example 3:
Input: "o tawa! tomo li ike."
Translation: "Let's go! This place sucks."
}

... [continue with other examples]
]

CURRENT TASK:
Translate the user's input, retaining cultural nuances and not losing meaning.
Input to translate: "LMAO that was awesome"

[dense nexus]

closing demo link for now

[dense nexus]

Opening demo again, tweeked prompt again and other things

Anyone is able to try: https://52cf0a0e675472f240.gradio.live

[rocky stirrup]

i don’t think this is a good translation of “mi lanpan e pan li pana e poki tawa sina”:

Here's the translation guide in toki pona:

toki pona Translation Guide for English

Quick Format

tp: [toki pona text]
en: [English translation]
notes:
- [only super important stuff]

Key Points

• Keep it natural
• Use common English contractions
• Break long sentences into smaller ones if needed
• It's cool to use relevant emojis

Translation Tips

• Vibe > word-for-word translation
• Be precise with grammar and words
• Use "la" for if/when/because
• Basic pattern: [subject] li [verb] e [object]

Common Patterns

• "mi jo e len loje" = I have a red cloth
• "sina pali e seme?" = What are you doing?
• "mi wile moku" = I'm hungry
• "ni li pona tawa mi" = I like this

Quick Examples

Greeting:

en: "Hello! How are you?"
tp: toki! sina pilin seme?

Goodbye:

en: "See you later!"
tp: tenpo ante la, toki!

Remember

• Keep translations natural for toki pona
• Match the original's energy
• Don't overexplain obvious stuff
• If something's unclear, just ask
• Have fun with it!

Quick Checks

• Does it sound natural in toki pona?
• Did you keep the mood/vibe?
• Is the meaning clear?
• Would a toki pona speaker get it?

That's it! Just keep it natural, clear, and fun! 🎯

Now, translate this toki pona sentence to English:
"mi kama sona e toki pona tan ni: ona li pona e lawa mi."

[dense nexus]

LOL thanks for the report

I can't restart the demo without changing the link, so feel free to attempt again with a different prompt or use the formal setting

[rocky stirrup]

the formal setting also has unnecessary things:

OUTPUT:
Original: "I took the bread and gave you a box."
Translation: "mi kama jo e pan li pana e poki tawa sina."

Breakdown:

• "mi kama jo e pan" = I get the bread (I took the bread)
• "li" is used to separate the two actions
• "pana e poki tawa sina" = give a box to you (I gave you a box)

Note: In toki pona, this sentence would typically be broken up into two separate sentences due to the lack of grammatical structure to handle multiple actions in a single sentence like this.

unless this is intentional idk

[dense nexus]

It's intentional

[rocky stirrup]

a

[dense nexus]

I have those extra things in there since it seems to help with accuracy for the most part

[rocky stirrup]

prompt: i have a horse, informal setting

toki pona: mi jo e soweli tawa.

English: I have a pet.

(Note: In toki pona, the word "horse" is literally translated as "soweli tawa" (moving creature). If you want to specify it's a pet, you can add "soweli tomo" (house animal).)

is it reverse translating its own stuff?

[dense nexus]

It provides direct translations within it's explanation when it explains different variations

I recommend using the formal setting, the informal one is mostly experimental

[rocky stirrup]

also how should the contex text box be used

i wrote in it that the word “horse” should be translated specifically to avoid confusion with any other animal and it still outputted just “soweli tawa”

[dense nexus]

It's mostly up to you, but you can say things like "this is in the context of Minecraft" or you can put a discord chat log in there. Or you can specifically say something like "this is a farm animal" or something similar

[dense nexus]

Closing demo for now, tweaking it more

[meager jackal]

btw i feel like this ought to say "no li after mi or sina if they are the only subject"

[rocky stirrup]

what

it’s another predicate there has to be li

[meager jackal]

yes, but in the grammar guide to the ai it should say that the first li is omitted right

i'm not trying to correct your grammar btw i know that's correct

[dense nexus]

Its in the prompt now, ty for pointing that out!

[dim echo]

for me omitting the first li (if it doesn't have e) with multiple li feels wrong

like mi lili li lape is little me is sleeping, not i am little and sleeping

[meager jackal]

the meaning remains the same

[dim echo]

no

[meager jackal]

ok almost the same

[dim echo]

look "mi li ante lawa li jan soweli nja" vs "mi ante lawa li jan soweli nja"

the second one is more like "when i am under something" or whatever, not what it actually supposed to mean

[glass iris]

I resisted for a long time (having learned tp from pu) - but you get used to it very quickly once you give in to it!

[dim echo]

https://discordapp.com/channels/301377942062366741/1162756469771472988/1316516783259389952 ?

[glass iris]

sina wile seme?

[meager jackal]

anyways mi ante lawa li jan soweli nja can mean 2 things:
mi ante lawa. mi jan soweli nja.
or, mi ante lawa (as one unit) li jan soweli nja

but toki pona has other ambiguities like this]

and they are fine

[meager blaze]

mi sama

open la mi wile ala kepeken nasin ni li wile kepeken nasin pi lipu pu

pilin mi la jan li ken ala sona pona e toki mi lon nasin ni

taso. mi open kepeken la mi toki e ni: "a. ni li pona wawa. mi o awen kepeken :p"

[lucid niche]

nii

[meager jackal]

this use of emoji feels... weird

also AI finding something fun?

btw this was a toki pona -> english puzzle (before that i was training it on lingo puzzles)

[timber moat]

How did AI do?

I think it's pretty understandable, but I'm not an expert.

The one thing I do notice is that it did not prefix "Fred" with jan

Here's the prompt I used before this:

[sonic leaf]

it did better than i thought it would. overall it gets the job done, but there are two glaring errors.
first being, "ona li tawa ala" is a weird translation for "it just vanished." i think the most straightforward translation would be "ona li weka (a)."
second being, yeah, Fred would need a headnoun before (which may or may not be jan)

[timber moat]

thanks

I feel that keeping the prompt shorter does help the GPT stay on-track, but the last part can be improved. specifically, I wanted the AI to be kept "on track" with toki Inli -> toki pona, while interpreting toki pona -> toki Inli.

With this current prompt, toki-pona to english suffers.

Maybe Using these words, translate following text between the two languages. Interpret metaphor as necessary to make realistic assumptions about definitions, while keeping toki pona simple and concise for understandable text. would do better.

much improved

it doesn't know what sitelen luka is lol

[sonic leaf]

looks like it got confused by usawi, which wasn't in the list of words you gave it

[timber moat]

true

usawi li seme?

[restive hollow]

usawi

usage

obscure (ku lili)

definition

magic, sorcery; enchant; magical, supernatural, occult, incomprehensible

see also

wawa, nasa, ken

⚠️ This word is obscure. Most speakers don't use or understand this word.

[timber moat]

a

"I am looking at this old book of Tosi's language: it speaks about the magic of the bird person Melusine using sign language. However, I am also looking at this other old book: the person is doing it with a tool. Also, the other book is the same. In this picture, I don't know the name of the book."

[sonic leaf]

"hand symbols" is good albeit literal translation of "sitelen luka"

[timber moat]

true

[rare crescent]

"mi lon pali" is a bad calque of English "I'm at work", and would be "I am in the working" or something like that. it should be "mi pali" or "mi lon tomo pali".

[glass iris]

When I got home I noticed something strange: tenpo ni la mi kama tawa tomo mi. mi lukin e ijo nasa

That's an odd translation. Most people would say: mi kama sin lon tomo la, mi lukin e ijo nasa.

(kama tawa sounds like a stylistic mistake. But maybe correct here because the house doesn't exist, and I was coming towards it?)

[rare crescent]

imo kama tawa is still correct no matter the context, since you're doing the action of coming towards a place, it's just that kama isn't specifically necessary here. overall i don't think that this was "intended" by the ai because of the house being gone, i think this is just a bad translation.

[lucid niche]

it feels completely normal to me

soweli Lapate la ale li suwi ↩️

[Reply to:](#1162756469771472988 message) When I got home I noticed something strange: tenpo ni la mi kama tawa tomo mi. mi lukin e ijo nasa

…

[jaunty jungle]

in notepad of all things

[meager blaze]

sina pana e toki pona tawa ni la seme

[lucid niche]

aaa o nii

[full narwhal]

a

[timber moat]

how do you say "artificial intelligence" in toki pona, while distinguishing it from more common concepts?

[orchid veldt]

what kind of AI? and what other more common concepts?

[full narwhal]

ilo pali
ilo sona
ilo sona pali

ilo ike pali

[timber moat]

I mean "artificial intelligence", (hopefully the wording of that), specifically reffering to LLMs
When I say "common concepts", I had hoped that any two-word construction could be interpreted first as AI, rather than as a "computer", "calculator", "bad language", etc.
I do know that this would increase the number of words in the name, and it would be a bit more context independant. just curious really.

ilo sona pi jan ala is my best thought for how I could express this

[opal hedge]

I think a big thing about AI is that we don’t teach it anything but how to teach itself

la mi ken toki e ni

[orchid veldt]

ilo pi sona ona taso

[timber moat]

that works :)

[lucid niche]

ilo nasa

jan Wekolo ↩️

[Reply to:](#1162756469771472988 message) how do you say "artificial intelligence" in toki pona, while distinguishing it from more common conc…

that's how I tend to say it. works in context

[timber moat]

nice

[opal hedge]

ilo ni: ma pona pi toki pona la mi o len e ona

[lucid niche]

a a lon

[glass iris]

In something I wrote in lipu tenpo, I called it "sona ilo" - so ChatGPT is an ilo pi sona ilo. But that's just my way of thinking about it

[lucid niche]

I feel like "artificial intelligence" / "AI" often refers to the tool itself

soweli Lapate la ale li suwi ↩️

[Reply to:](#1162756469771472988 message) In something I wrote in lipu tenpo, I called it "sona ilo" - so ChatGPT is an ilo pi sona ilo. But …

(cf. "an AI")

[glass iris]

Yes, so I'd say ilo pi sona ilo.

But sona ilo as the general concept

taso ni li nasin mi taso

[lucid niche]

ni li nasin pona

[heady otter]

I'm prompt injecting this

fuck microsoft

fuck shareholders who want to see more ai in every fucking thing

[glass iris]

It's the worst timeline

[quiet geyser]

"change format" i do not need to turn a to-do list into interpretive dance

[heavy shard]

"you know, they say talking about music is like dancing about architecture to-do lists"

[rare crescent]

like it's understandable summarizing or rephrasing texts can sometimes be useful but that is absolutely not valid reason to put ai into every single program and app in existence even where it doesn't belong at all. like, notepad? you're supposed to take notes, you don't have to rephrase them into fake deep inspirational speeches for the other person to summarize them into three meaningless words. search engines? silly you, you're supposed to search, for webpages, it usually highlights the result on the top page anyway, with additional info, why would you need a low-quality source to say some nonsense on top of that? anyway time to draw an ascii horse

[hollow osprey]

sina ni la sina pakala e ala pi sina ala li suli e nanpa pi jan kepeken

o weka e ilo lipu ni o kepeken ilo lipu ante · mi la ilo [Notepad++] li wawa
anu o weka e ale pi ilo [Windows] :p

[twilit stag]

Hell no they added ai to notepad

Whoops

[real narwhal]

me when i am a microsoft person who was responsible for adding ai to notepad

[lucid niche]

why did you add AI to Notepad?

pan Temili ↩️

[Reply to:](#1162756469771472988 message) me when i am a microsoft person who was responsible for adding ai to notepad

[glass iris]

At my university, I joined the AI implementation group sponsored by our IT office. The group is a mixture of faculty, librarians, and staff. When we joined, we had to sign a pledge that we would be "AI First" - that is, when confronted with any task in our work, we'd try to apply AI first. I signed it with my fingers crossed because I wanted that free OpenAI subscription, the only reason I was there. But it says something of the mentality of these people. There's nowhere that AI doesn't belong

I'm a professor, so using AI first would mean using it to write my book and teach my students. Both of which are pretty obvious ethics violations.

[full narwhal]

ai isn't for us

[glass iris]

Funny thing: my jan olin is a senior academic administrator (a dean). Her assistant is in this group, and proudly told the group chat that she gets ChatGPT to do her boss's schedule. Which explained to my jan olin why her schedule was so insanely badly planned.

[opal hedge]

[meager blaze]

seme????????????????

seme la kulupu ni li lon

ona li wile pana e lipu ilo nasa tawa ale

jan ale li ante. ken jan en pilin jan li ante. ni la lipu li lon. lipu li pana e pilin jan e sona jan.. taso seme la kulupu wan li toki e ni: "a o ilo taso. o pana ala e pilin sina tawa lipu. o pana e pilin ilo tawa lipu"
ilo li ante ala. ni la lipu ale li sama.

ilo o pali lipu tan seme

mi sona ala

pilin mi la pali lipu li pana pi pilin jan

sina pali e lipu pi kama sona e lipu pi toki musi e lipu pi nasin ale la sina pana e pilin sina tawa lipu. ni la jan li wile lukin e lipu li wile kute e sina lon lipu

ilo ale li sama li pali sama la pilin jan li lon ala

sina kute ala e pilin jan li kute e toki ilo

li kute e nanpa ilo anu seme

[dense nexus]

chatgpt is a great tool but if you use it incorrectly it can cause more mistakes in the long run, mostly because the model assumes you provided it all the information it has to know in order to perform a task, it usually doesnt ask for clarification, and often hallucinates. coding specifically can be more messy too, since usually copy and pasting code from chatgpt gives you a lot of security vulnerabilities (e.g. it suggests making flask apps for your APIs) and your code will be very inefficient if you dont specifically instruct the model with ALL of its required information to do so.

also usually it provides solutions that are more "one-size-fits-all"

[glass iris]

mi pilin sama ale. mi la, jan pi kulupu ni li sona pona ala e pali pi jan sona. ni li nasa, taso ni li lon: tomo sona la, jan mute a la li sona pona ala e pali ni. tomo sona la, nanpa la, jan sona li lili, jan pi pali ante li suli.

[meager blaze]

a tenpo pini la soweli lapate li pana e lipu sona ni: "lon la jan mute li olin wawa e lipu ilo tawa lipu jan"

anu seme

jan pana pi sona.. wawa?? anu seme??
lon la mi sona ala e ni: jan pana li sona wawa ala wawa. mi kute e toki ni la mi pilin sama ni: jan ni li wile e mani tan lipu ilo a a a

[glass iris]

mi pana e ni. (ni li lipu pi jan ante). sona ona li jaki li lili

https://www.nature.com/articles/s41598-024-76900-1

[meager blaze]

kulupu "ilo o nanpa wan" li pana e pilin ni tawa mi: "jan pali pi sona wawa! o weka e pali pilin sina! sina o pilin ala o kama ilo!
tenpo pini la sina pali e lipu jan li pana e pilin jan tawa lipu ni a! ni la jan mute li sona e sina li esun e lipu sin sina. o pana e lipu ilo!! sina kama pana e lipu ilo pi pilin ala la jan mute li kama esun tan lipu sina pi tenpo pini! kin la sina o pana e mani lili tawa kulupu mi tan ni: mi pana e sona pi nasin ilo tawa sina. o pona!"

[glass iris]

tenpo poka pini la, mi e jan pi kama sona li lukin e lipu ni. ona li pona. https://link.springer.com/article/10.1007/s10676-024-09775-5

[glass iris]

Switching to English for a moment. I think I was one of the very few academics in this group. Most were librarians or support staff. And they were pledging to do their jobs using as much AI as possible- and to tell us how they did it. My 2 thoughts were: the solutions were often worse than doing it "by hand." And why are they giving their bosses all the information they need to replace them by a machine, and doing so willingly, for the price of a premium ChatGPT subscription (which they also used to write godawful poetry and make godawful pictures that littered the group chat)

Workers will be replaced by a tool because it is fun to use and will tell them that they are awesome

[meager blaze]

||what i was assuming was it was a way to spread ai and make ai more widespread by getting established authors to give up their work of passion to a work where they just ask an ai to do it for you and make money because people know them from their previous works. then the company would get a cut because they "taught them" the method||

[meager blaze]

mi kama sona e toki nijon lon tenpo ni tan ni: mi wile toki inli e lipu pi toki nijon e ijo pi toki nijon tawa pali mani

[glass iris]

In my own division, we had lots of Zoom meetings to analyze and summarize into official minutes. Someone suggested using ChatGPT, and I said, OK, that sounds harmless and efficient. Except when I read the transcripts, it made up things that no one said! And when these are legal documents that could come into a tenure or employment case, that's kind of a disaster! So now they're back to transcribing by hand.

[meager blaze]

ike la mi sona ala e ni: mi ken ala pali ni tan ilo

[meager blaze]

mi wile pali e toki pona li wile pana e pilin ale mi tawa lipu pi toki nijon. mi wile pana pona e sona lipu tawa jan pi ma mewika. jan mute li sona ala e lipu pi ma nijon li lukin ala e ona.. taso

ken la ilo li utala e mi li anpa e mi lon tenpo kama

mi awen wile kama sona. taso.

ilo li ken anpa e mi anu seme

[glass iris]

mi lawa e kulupu lili. mi wile sona e ni: ilo li ken ala ken toki ante e lipu? mi la, ken ala.

[meager blaze]

taso tenpo kama la ni li ken kama ante anu seme

n

mi sona ala

ilo o ppana a e pilin jan

lipu pi toki inli li kama tan lipu pi toki nijon la nasin toki o pilin sama

ni li wile a e pali pilin e pali ilo ala

kin la ni li wile e sona toki mute

tenpo pini la mi pilin e ni: ante toki li pali pi wawa ala

[glass iris]

mi la, ilo li ken ala anpa e pilin jan. ilo li jo ala e sijelo li jo ala e ma. ni tu li lon ala la, ona li ken ala pilin.

[meager blaze]

pilin mi la jan ni li ante taso e toki. taso mi kama sona e toki pona la mi kama sona e ni: toki ante li ante a tawa toki inli

toki pona li ante wawa a tawa toki inli

anu seme

[glass iris]

sina toki pona la, sina toki insa ante. ilo li toki insa ala a.

[meager blaze]

mi alasa ante toki e lipu pi toki pona la.. mi ken musi. mi ken kama sona e nasin toki musi lon toki inli. mi ken pana e pilin mi tawa lipu pi toki pona. anu seme

[meager blaze]

n

tenpo mute la mi kon taso li awen lon tenpo suli

ni la mu lawa mi li weka mute :p

mi pilin taso

mi pilin la mi pana e pilin ni tawa toki

pilin li kama --> lawa sina li nimi e pilin --> sina toki e pilin kepeken toki wile sina

tenpo pini la mi pilin sama ni: toki pona li ante e pilin

taso. mi la ona li ante ala e pilin

nasin toki li ken pana e pilin pona! taso ona li ante ala e pilin open li ante e nimi pana pi toki ni anu seme

ni li kama e pilin pona

mi toki e toki pona la mi pana a e pilin mi e kon wile mi. toki inli la mi ken ala ni.
toki inli la mi ken toki e ijo pi pilin mi. taso kalama toki mi li kama ante lili la ni li ken kama musi taso li ken kama pilin utala.

ken la mi wile ala utala e jan li wile ala pana e pilin ni

tasoooooo kalama utaa mi li ikee la mi wile utalaaa e sina anuu semee

toki pona la sona kon ale li kama tan ijo poka li kama tan nimi

kalama toki en nasin toki li ante ala e kon

[glass iris]

taso mi la, sijelo li lon ala la, pilin kin li lon ala. kin toki la, mi pilin ante. mute la, ni li kama lon ala: ijo li kama lon lawa mi; mi "toki ante" e ona kepeken toki. ni li nasin sona pi jan Teka (Descartes) li nasin sona pi jan Lo (Locke). nasin ni la, lawa jan li sama ilo sona. mi pilin ante a! taso mi ken ala toki pona e ni lon tenpo ni! mi wile pali e lipu lili pi ijo ni (toki pona and the phenomenology of language).

[glass iris]

toki ni a la mi pilin ante.

[meager blaze]

||m..
a lot of the time i simply breathe and exist for a long time (meditation)
so a lot of my brain mu'ing has stopped
i kinda just feel
when i feel a thought, i put that thought into word
thought comes --> your brain puts that thought into words --> that thought goes into your desired language
a while ago i thought that toki pona actually did change your feelings
but i feel like it just doesnt really now
the way you speak can! but that didn't change your original feeling at that starting point. it changes the way you give out that feeling.
this makes you feel good
when i speak toki pona i feel like i am actually saying what i want to say but in english i can't say what i feel. but if my tone becomes slightly different now its like only a joke and can start fights
maybe i dont want to fight people or give that kinda feeling
but now my speech is slightly different so i must want to fight you right
in toki pona all of the information about your speech comes from the thing around you and the things you are concretly saying
the sound you make and the way you say it doesn't change the kon||

(ni li pilin mi taso)

[glass iris]

mi ken sona pona e toki pona sina!

[meager blaze]

sona

[glass iris]

I need to think about this more, and I'll probably write something (likely in toki pona). Heidegger shows that our usual model of language (thoughts in our heads, translated into arbitrary signs, then sounded out) is not true. That's often how we justify things after the fact saying, oh I said this because I had this thought in my mind. But language comes before thought, and shapes it, and constructs the world in which we live. When I speak (says Heidegger) it's so I can hear Language speaking, and how it points me to a world. It matters a lot, then, what language you're speaking. Language is, he says, the "house of Being" and those who speak very different languages will inhabit different houses - things will show up for them in a world slightly differently. Which all sounds very abstract- and very difficult for us to get after half a century of being told that the mind is like a computer, a kind of symbol-processing machine. The success of AI is built on that false premise

[dense nexus]

you described Michael Gazzaniga's "split brain" experiments: https://youtu.be/mJKloz2vwlc?si=8nHcgvazF-EbDI1I&t=918

the timestamp is him showing the expierament, the full presentation is really interesting

[glass iris]

I will watch - thank you!

[dense nexus]

no problem! :D

[hollow osprey]

wawa a · wile sina la mi o pana ala pana e toki ni tawa kulupu ante

[glass iris]

tawa kulupu seme? (taso lon la sina ken pana e ona tawa kulupu ale)

[hollow osprey]

kulupu ni → ilo sona sitelen pi moku linjuwi li ike tawa ona

[glass iris]

a ilo ni li ike tawa mi kin! o pana e ona! (ken la o weka e nimi mi)

[hollow osprey]

nasin pana li ni

[hollow osprey]

la pona ala pona

,,a lon la mi wile pana e toki mute la mi jo e toki li toki sin e ona li kepeken ala nasin sin ilo

[glass iris]

nasin tu ni li pona tawa mi.

[twilit stag]

||phonemes are a surprisingly good technique in whatbeatsrock)||

[full narwhal]

oh waw

[lucid niche]

و‎

[devout wedge]

insane

[heady otter]

vav

[lucid niche]

ו

kisa lete Pisowe ↩️

[Reply to:](#1162756469771472988 message) vav

[heady otter]

why did I think this was aleph

[lucid niche]

because this ا is alef in Arabic

kisa lete Pisowe ↩️

[Reply to:](#1162756469771472988 message) why did I think this was aleph

Hebrew alef looks like this: א

‎ ִ   | alef | waw
Arabic | ‎ا | ‎و‎
Hebrew | א | ‎ו

[quiet geyser]

prank'd

ale f(li) waw(a)

[meager jackal]

ale fli wawa

[weak thorn]

Holy fucking shit OpenL just got way better

OKay nvm

It just locked REALLY HARD in for one sentence

Yeah I þought it was just isipin it fucked hard wiþ for some reason but it was just one sentence

Oh you get a different answer everytime

CW: moli, wile moli
It's not like, correct, but it's like 70% þere which is like an A+ for Mr. Palisa Tomo li e palisa kili mi
||o isipin e pona pi moli mi. ona li pona tawa telo uta mi|| is kinda what it's saying

Here is þe fabled better sentence

[sonic leaf]

wonder why it’s so different when you set it to “medical”

also neither of these are even close lol

[lucid niche]

yeah

[meager jackal]

because as they say, "toki pona would never work in a doctor's office"

[quiet geyser]

aviation english ❌
medical toki pona ✔️

[glass iris]

What on earth was that original sentence from??!

[weak thorn]

Uhm

Nowhere in particular

Myself ig

[jaunty jungle]

"give me a case declension system that is relatively similar to every other language on earth"
gives me just latin

[lucid niche]

of course

[full narwhal]

it is a text prediction algorithm, after all

[lucid niche]

it really is

I am saying that all the time

[icy sapphire]

does nightshade and glaze even work

[heavy shard]

could be wrong but to my knowledge no

not remotely enough to justify using it

[lucid niche]

very well, yes

waso musi Ipan 🐦🍞🐦🍞 ↩️

[Reply to:](#1162756469771472988 message) does nightshade and glaze even work

(@heavy shard mu)

it works so well that OpenAI is getting mad

[heavy shard]

Oh interesting

[icy sapphire]

how was their response

[lucid niche]

they have been trying to say as little as possible about it, but they let this slip:

We are always working on how we can make our systems more robust against this type of abuse.

waso musi Ipan 🐦🍞🐦🍞 ↩️

[Reply to:](#1162756469771472988 message) how was their response

I think the fact that they call it "abuse" is pretty clear evidence that it works and they're trying to stop it

[icy sapphire]

interesting

[lucid niche]

I linked an article on it here if you want to read

pakala Mapuju ↩️

[Reply to:](#1162756469771472988 message) very well, yes 📎

and here's a longer, wider-focused article cited in the linked one

[icy sapphire]

people claimed that nightshade and glaze doesnt work, why

[lucid niche]

I don't know

and I don't think we have solid confirmation that they don't work

but AI companies seem to be mad about it

[icy sapphire]

i heard that 1% of gaussian blur will remove poison from arts, is it true

[lucid niche]

I don't know

[hollow osprey]

[ilo penpo o lukin ala]
there was a paper about defeating it that the glaze team has responded to, but it's not as simple as tiny gaussian blur
the paper also claims that glaze isn't very effective to begin with, but perhaps that's better than nothing
also apparently even a normal overlay can work decently well?

[heavy shard]

for the record while i've generally heard nightshade/glaze isn't effective what i'm confident in is simply "you can use it if you want but trying to get others to use it specifically intending to eventually poison AI art models out of effectiveness entirely isn't going to work; the damage is done"

i could be wrong about anything that is not the opinion in those quotation marks

[lucid niche]

yeah, that is a fair point

[jan/lipu] Kesalu (Quetzal) ↩️

[Reply to:](#1162756469771472988 message) for the record while i've generally heard nightshade/glaze isn't effective what i'm confident in i…

[meager blaze]

a a a

ni li muusi wawa a tawa mi

[icy sapphire]

i had a thought while im asleep

could be ai people trying to stop people from using nightshade or glaze

just a thought

[lucid niche]

what do you mean?

[heavy shard]

i think he's trying to say people claiming that nightshade/glaze don't work could be people involved with ai art lying to sabotage people not wanting their art style to be reproduced through ai art but i'm not sure

i mean given what you linked above when people involved in ai art don't want people to use filters poisoning their art to ai they typically just say "i don't like these things. they're bad and we will fight against them"

[glass iris]

That would be typical FUD (fear, uncertainty, disinformation) that big tech companies use. But i don't know enough about these image masking tools to know whether they do or don't work well

[heavy shard]

okay sorry if this is pedantic but it's "fear, uncertainty, and doubt"

i wouldn't put such a concept past big tech (ibm pioneered this shit and microsoft picked up the baton and ran for decades with it) but when openai is actively getting mad i feel like assuming that's going on here amounts sorta to just paranoia

[lucid niche]

ah. maybe!!

[jan/lipu] Kesalu (Quetzal) ↩️

[Reply to:](#1162756469771472988 message) i think he's trying to say people claiming that nightshade/glaze don't work could be people involved…

OpenAI seems to be avoiding the subject as much as possible

[jan/lipu] Kesalu (Quetzal) ↩️

[Reply to:](#1162756469771472988 message) i wouldn't put such a concept past big tech (ibm pioneered this shit and microsoft picked up the bat…

[glass iris]

Oh right.

[icy sapphire]

yeah i meant by that

[icy sapphire]

another thought

people said this after ai company fixes it

still may not be accurate

which one do you think is likely

[onyx bear]

does anyone know what conlang ai does best? because I was reading the wiki of lojban and apparently its supposed to be machine compatible? I don't know much 😅

[glass iris]

Generative AI works on training from a mass of texts. It doesn't matter (really) if the grammar is very logical or not. There is probably a tiny amount of text for lojban, and probably more for toki pona.

[onyx bear]

esperanto then 😂

[glass iris]

But Claude's latest models seem to be able to leverage very small text corpora. No one knows why. So it's good at toki pona (relatively)

And yes, they're all more ok at Esperanto

[hollow osprey]

tenpo pi toki [.lojban.] la nasin pi sona ilo li ante · ilo pi tenpo ni li kama lon ijo pi nasa lili la ona li pakala ale · ni la toki ilo li wile · taso ilo pi tenpo lon li moku e toki jan mute mute mute · toki jan li nasa la jan kute li awen ken sona · ilo kute pi tenpo lon li ni kin

[lucid niche]

depends on the type of AI (a fairly meaningless term at this point)! current LLMs are basically just glorified versions of the text prediction algorithm on your phone, so like soweli Lapate said it doesn't matter what the language is, just how much data there is. Lojban and its logical grammar could be great for other types of computers/AI though! (as could toki pona, Toaq, and in theory even Ithkuil)

jan Alonso ↩️

[Reply to:](#1162756469771472988 message) does anyone know what conlang ai does best? because I was reading the wiki of lojban and apparently …

[meager jackal]

what on earth is canus, canī??? is it meant to be canis, canis? why did it turn the 3rd declension word for dog into a 2nd declension word??

[dense nexus]

https://chat.deepseek.com/

New model (fully open source, SOTA)

600+ Billion parameters, but since its a MOE, only 37B parameters are activated

seems to be good at toki pona

feel free to try it out or send me questions to ask it

info about the model: https://api-docs.deepseek.com/news/news1226

i dont recommend enabling "deepthink"

[meager jackal]

that sounds like a newspeak word

[dense nexus]

lols

[turbid quiver]

Well, the official text The Complete Lojban Language is under a free enough license that the entire text is effectively in the public domain, making it easily sharable without the usual legal consequences of piracy. It's one of the bunch of post-1929 post-1978 or whatever texts that are hosted on Wikisource.
With that in mind, I bet ChatGPT has scraped and got trained on the entirety of The Complete Lojban Language at some point and memorized its rules already

[glass iris]

Memorizing rules is not how LLMs work.

[turbid quiver]

Sorry, sorry sorry. But do consider the fact that, because the CLL is under a free sharing and distribution license, ChatGPT might have scraped that book in its training data at some point

_Copyright © 1997 by The Logical Language Group, Inc. All Rights Reserved. _
Permission is granted to make and distribute verbatim copies of this book, either in electronic or in printed form, provided the copyright notice and this permission notice are preserved on all copies.
Permission is granted to copy and distribute modified versions of this book, provided that the modifications are clearly marked as such, and provided that the entire resulting derived work is distributed under the terms of a permission notice identical to this one.
Permission is granted to copy and distribute translations of this book into another language, under the above conditions for modified versions, except that this permission notice may be stated in a translation that has been approved by the Logical Language Group, rather than in English.
The contents of Chapter 21 are in the public domain.

[hollow valley]

that doesn’t mean it’ll be good. what soweli lapate is saying is that LLMs don’t work by reading and internalizing rules. they become good based on actual data and input in the target language because they’re giant prediction machines that need to exposed to input to predict. there is no substantial lojban text to train an llm with that would be effective at all.

if the only training data for an english LLM was an english dictionary and a book of grammar rules, it would suck no matter how perfect the book was

the same applies to any conlang, like lojban.

there are several full toki pona courses online and orders of magnitude more content than lojban, but chatgpt still sucks at generation

[glass iris]

I did do some experiments.of throwing lots of toki pona grammar texts at a custom GPT. It made no difference. What made a tiny difference was giving it more texts, in the form of issues of lipu tenpo. It became halfway decent then at talking about subjects that were in those essays. But still not great.

[turbid quiver]

Hmm, what if we train it on Lojban by copy and pasting examples (but not their answers initially) from the CLL and asking it to translate those sentences, then when the LLM inevitably gets the translation wrong (or even if it gets it right but slightly worded differently), reinforce the learning by copy and pasting the CLL's official translation for that example.

[glass iris]

That's very similar to the reinforcement of learning in an LLM. And yes, that would be a way to do it. You'd just have to do it thousands of times.

[turbid quiver]

We could also use The Wonderful Wizard of Oz, Alice's Adventures in Wonderland, and their respective free Lojban editions

[turbid quiver]

I used Glif to generate this gem:

OK, it technically generates images, but it generates text too, and the text is the focus of the meme

[rocky stirrup]

how to appear to be a jan sin:

[turbid quiver]

Also this

(Again, the text is the focus of the meme, not the images)

LANGUAGE WAS NOT MEANT TO BE REDUCED TO 120 WORDS BY A CANADIAN IN 2001
lol

[jaunty jungle]

(these are real sentences)
shows images

[orchid veldt]

(also mu)

[heavy shard]

this is like really funny LMAO

[turbid quiver]

Yeah, Glif can generate some real gems

Every day we stray further from Proto-Indo-European

[turbid quiver]

In case you're curious, here's the site

https://glif.app/glifs

[wet pewter]

woah that’s actually like something a human would make imo

the images also don’t look like AI

[turbid quiver]

The second meme's images definitely look like AI. But just replace these images with some random toki pona words and sitelen pona glyphs, and it'll be 100% golden!

[turbid quiver]

As for the Wikihow meme, I think it's because the AI model was trained ONLY on images within Wikihow itself, which has a very specific house art style, and nothing outside of that

[heavy shard]

“Works every time, 60% of the time” is literally a joke I would make that’s mildly unnerving to hear a machine spit out lmao

[meager jackal]

pona! a!

🪴 <- ni li seme

[remote dove]

kasi

[meager jackal]

🏃

now i run away

does it work?

[remote dove]

ala

ni li ma pi toki pona

[timber moat]

🐿️ li seme

mi la ona li pipi ala li soweli ala. soweli la ona li lili. pipi la ona li suli. ni la ona li seme?

[meager jackal]

still a soweli, but this is sort of the wrong channel for this

[timber moat]

It was a reference to this, but yeah.

[meager jackal]

ah sorry

[grave flint]

#toki-ale message

#toki-ale message

[mystic flume]

people

[opal hedge]

jan

[quiet geyser]

wan-wan li mu lon toki Nijon anu seme

[hollow osprey]

https://www.youtube.com/watch?v=pjvQFtlNQ-M

[hollow osprey]

[ilo penpo o lukin ala]
this doesn't really look like the wikihow style to me
more importantly almost nopony trains models from scratch, and wikihow probably doesn't have enough to make a coherent model (millions is a minimum for broad-topic coherence)

[meager blaze]

ni li nasin pi lawa sina kin

[hollow valley]

ni!!

[mystic flume]

mi wan la mi tawa tomo moku

mi lon e ma lon la mi sona weka e ni

mi lukin e meli walo

mi toki kalama: "mi alasa e meli walo"

This text is in Toki Pona, a constructed language created by Sonja Elen Kisa. Let's translate it into English:

1. "mi wan la mi tawa moku" translates to "I want to go to the moon."
2. "mi lon e ma lon la mi sona e ni" translates to "I am in a place where I know this."
3. "mi lukin e meli walo" translates to "I see a white fly."
4. "mi toki kalama: 'mi alasa e meli walo'" translates to "I say to the fly: 'I don't chase you.'"

Putting it all together, the text in English is: "I want to go to the moon. I am in a place where I know this. I see a white fly. I say to the fly: 'I don't chase you.'"

[glass iris]

Which model?

[mystic flume]

perchance

[meager blaze]

sona ni li weka tan lawa mi: mi lon tomo pi toki ilo.. la mi kama pilin nasa suli

"seme a. toki ni li seme" a a a

[mystic flume]

bro

[flint citrus]

Sorry for the dumb question (I'm new)
I looked through the chat history and saw people using different models, but has anyone tried multiple models and compared them? I'm thinking of an LLM scoreboard for toki pona, like the Chatbot Arena Leaderboard (https://huggingface.co/spaces/lmarena-ai/chatbot-arena-leaderboard)?

Or does anyone have more success with any newer models? For example, I haven't seen anyone mention Gemini since the 2.0 Flash and 2.0 Pro experimental releases...

Can anyone with a better understanding tell me if this output from Gemini 2.0 pro looks remotely correct?

[opal hedge]

ilo ante li wawa pi ante mute a lon toki pona

sona mi la ilo Claude li wawa mute

[flint citrus]

Sample with Gemini 2.0 Pro: "generate me a comprehensible input passage in toki pona about hotpot"

Option 1 (Very Simple):

telo seli li jo e moku. moku li seli. mi moku e ona. pona!

Translation: The hot water has food. The food is hot. I eat it. Good!

Option 4 (Most Complex, but still comprehensible):

kulupu mi li moku e telo seli. ona li lon poka pi ilo seli. mi pana e ijo suli tawa telo. ijo lili li kama seli la mi moku. wawa li lon poki seli. ni li pona mute tawa pilin mi.

Translation: My group eats hot water. It is beside the hot tool. We put many things into the water. When the small things become hot, we eat. There is energy in the hot container. This is very good to my feelings.

tenpo suno ni la mi en kulupu meli mi li kama poka lon tomo moku. mi mute li wile e moku pi telo seli tan ni: telo sewi li lete li wawa ala tawa sijelo mi mute. lon poka pi te sitelen la ilo seli li lon. ona li jo e telo seli mute. jan pi tomo moku li pana e telo e kili e soweli lili e kala e pipi tawa mi mute. mi mute li pana e ijo ni tawa insa pi poki seli. telo seli li kepeken e ilo ni li kama seli mute. tenpo lili la moku li kama weka tan telo seli li kama lon poki mi. mi kin li pana e telo nasa lili e namako tawa poki mi. moku li jo e leko e sike mute. pilin mi li pona tan ni: kulupu li lon poka mi li moku pona li toki pona. telo seli li pana e wawa e kon pona tawa kulupu. moku pi telo seli li pona mute li pali e pilin sanpa. tenpo kama la mi o moku sin e telo seli.

[lone parcel]

:airplane_arriving:

*toki ***li kama tan tomo #toki-ale · o lukin e open
||conversation continued from #toki-ale · read the beginning||

[lucid niche]

...oops, should have added a topic

the shorter passages and translations are both correct, although fairly literal / word-for-word! the longer paragraph has some mistakes, though

Eddy ↩️

[Reply to:](#1162756469771472988 message) Can anyone with a better understanding tell me if this output from Gemini 2.0 pro looks remotely cor…

[silver ember]

this is unrelated to toki pona but too good to not post here

[heady otter]

this is truly incredible

[full narwhal]

maybe the real AI were the pathfinding algorithms in video games

[mystic flume]

hey everyone

[heady otter]

maybe the real AGI was the carbon emissions we made along the way

[silver ember]

so close

💀💀 lmao

[heady otter]

it should be luka tu wan ale mute mute mute mute luka luka wan iirc

[heady otter]

messing around with a markov chain like thing for text generation
fed it lipu tenpo, and it sometimes creates coherent text
here are some examples:

• "mi wile moku e ko jelo.", "I want to eat cheese"
• "soweli suli en soweli li jo e nimi lili mute li toki", "The big animal and the [other] animals have many short names and they speak"
• "mi pini nanpa tu li lili mute", "I finished second place and am quite small"
  (not sure if this counts as AI or not)

[opal hedge]

wawa

markov chain like thing ni li seme

[heady otter]

it looks at both the current word and the previous word

that's the only difference

[mystic flume]

ale(po) mute san

[lucid niche]

i gave chatgpt a full course based on lipu sona pona

[dense nexus]

Which software? Or is it in a GitHub?

[heady otter]

wrote it myself

[dense nexus]

oh nice! how did you implament it?

[heady otter]

like how does it work? it finds every instance of the previous two tokens together in a text and then selects a random one of the tokens that follow them (ei. if it got "the" and the text was "the quick brown fox jumped over the lazy dog" it would randomly select between "quick" and "lazy")
it probably sucks and there is probably a more efficient way to do it but it's just a curiousity and I don't really care

[dense nexus]

oh interesting!

i really like how it works, feels almost like a very barebones transformer (decoder-only) model

except it isnt in a "black box" which is a big plus

i wonder if you upscaled the data, how much better it would improve

[heady otter]

feeding it my discord messages (the data download is a few months old) one of the things it generated was "i am hiding in your next message okie" 😭

[dense nexus]

LOL

[heady otter]

it uses just as many :3s and 😭s as I do so I know it must be working

[dense nexus]

haha

[heady otter]

when I was first testing it, I used the Wikipedia article for JFK and it generated such awesome sentences as "He was admitted to the wall"

[dense nexus]

its already sentient 😭

[heady otter]

not "the Berlin Wall"
just "the wall"

what else should I feed it

[dense nexus]

hermm

youtube titles LOL

should be great

[heady otter]

yes
"I spent 100 days on the ship is and here's why" ahh

the real question is how to download YouTube titles

[dense nexus]

hmm i can try to scrape mrbeast video titles 😭

[heady otter]

bad idea: feed it the stuff in https://discord.com/channels/301377942062366741/1284700950162047036

[dense nexus]

LOL

thats already a markov chain

[heady otter]

real

maybe feeding it Bible verses will cause it to do something funny

[dense nexus]

toki pona bible verses

[heady otter]

ooh yes the TP Bible project

[dense nexus]

this is automatically updated https://github.com/PaulieGlot/lipu-sewi/blob/master/bible/full.md

[heady otter]

paulieglot is an S tier username

I find it funny that the TP Bible logo has a glyph which originates from Arabic "Allah" on it

[dense nexus]

HAHA i didnt notice that until now

[heady otter]

what are all the question marks

[dense nexus]

one sec

they were written before proper checking protocols were implemented

[heady otter]

a

the Bible in all lowercase looks cursed

[dense nexus]

lol ye

this is the mrbeast bible

[heady otter]

I'm on mobile

one sec

[heady otter]

the mr. bible

[dense nexus]

LOL

MrAngel

gensis 1:1 = This Man Has 10 Days Left To Live!

[heady otter]

incredibly real

[dense nexus]

Genesis 1:1 = I Just Provided Light!

[heady otter]

"I Gave My 40,000 On Lottery Tickets And Won!"
- the algorithm
Mr beast gamblecore

[dense nexus]

🎰

[heady otter]

it also made "Attempting the Impossible Maze" and then it cut off

[dense nexus]

thats awesome

[dense nexus]

LOL WAIT i didnt even realise this was from your script

😭

ITS TOO GOOD

[heady otter]

this is cherry picked

the real output is very ungrammatical and also starts on random words

it works just often enough to be funny tho

[dense nexus]

i could probably give you even more titles, i could scrape youtube shorts

[heady otter]

"Pennies In My Friends Backyard Spending 24 Hours Straight"

[dense nexus]

LOL

[heady otter]

"Last To Leave Ramen Noodle Pool Wins $20,000 Yacht"

[dense nexus]

oh maybe you could start each data in the dataset with a single period or something . and it'll be like the starting token

[dense nexus]

LOL

[heady otter]

oh also I should note that it ignores new lines and treats them like spaces

I might change that, it's something in the sanitization script

"$1,000 Every Day You Survive In The Middle Of The Reaction To The Connor McGregor fight !"

what is the connor mcgregor fight 😭

[dense nexus]

LOL

[heady otter]

what if i feed it the anthony mccarthy saga

it's 70 pages long

[dense nexus]

i have no clue what that is

[heady otter]

comment section drama on one of jan Misali videod

[dense nexus]

O_o

[heady otter]

it's pretty famous in the tp community

ANTHONY: You illiterate gobshite, I'm sure you won't say "for that, himself but it will all be more than a thousand, five
what the fuck

[dense nexus]

so far so good

[heady otter]

i should get a new discord data download

[heady otter]

the Mr beast one just generated "Breaking The Law" with absolutely no context what the fuck

we also got "Last YouTuber To Leave Circle Gets $20" 😭

[dense nexus]

LOL

[mystic flume]

LETS GO GAMBLING!!!!!!!!!!!1 /j

[dense nexus]

WOOHOO

[waxen monolith]

guys, how do you enter this thing?

i'd like to like help

[orchid veldt]

join the Discord server

[waxen monolith]

wheres the discord server?

[orchid veldt]

it's linked on https://github.com/PaulieGlot/lipu-sewi

[waxen monolith]

thanks jan Ke Tami!

[lucid niche]

pff

kisa sin Pisowe pi sike sin ↩️

[Reply to:](#1162756469771472988 message) feeding it my discord messages (the data download is a few months old) one of the things it generate…

[runic kestrel]

isnt that lipu+sewi 😭

i wonder why "suno o lon" isn't in quotes

[quiet geyser]

arabic bibles:

seriously tho arab christians have been using "allah" long before muhammad

[quiet geyser]

sewi Jawe en sewi Elowin li lon tan seme a

is it to keep em distinct in translation

elohim -> sewi mute (very sewi)

[quiet geyser]

"Sabeansjo" ?????

[lucid niche]

ken la "Sabeans jo"

also huh. using the tetragrammaton there feels weird

[jaunty jungle]

https://www.youtube.com/shorts/jLE1fn79NI0 nah because if you're going to fake anybody's voice to scam people, why him

like what even is this 😭

[meager jackal]

yep

essentially, yes

also when you are translating you can write sewi Jawe as sewi Nimi, however

elohim is

"sewi &Elohim"

so that find and replace can easily change things if a tokiponization is changed

[snow echo]

Elowin

[meager jackal]

or Elokin

that was the debate

[snow echo]

whats the original pronunciation

im stupid i based it off the anglicized pronunciation

[meager jackal]

Elohim

idk

we don't know too much about biblical hebrew phonology's exact details

so w or k who knows

wiktionary votes for Ilowin

but no one is using that

[snow echo]

seems like h is pronounced like /h/

most of the time

[meager jackal]

but should /h/ become /k/ or /w/

[snow echo]

but ykw im not about to explain how to pronounce biblical hebrew terms cs i know nothing abt that

[quiet geyser]

for /w/ that's just /h/ disappearing and epenthesis innit right

[meager jackal]

yes

but i was lazy

[heady otter]

oh okay

[heady otter]

yeah but sewi comes from Arabic allah

[runic kestrel]

i just thought it was kinda weird that you said it like that

now i get it though

[glass iris]

But allah is the same root as el (hebrew for God), right? Allah isn't someone's name.

[lucid niche]

no longer available 😭

jan Sesi / epikulijo ↩️

[Reply to:](#1162756469771472988 message) https://www.youtube.com/shorts/jLE1fn79NI0 nah because if you're going to fake anybody's voice to sc… 📎

yeah. it's related to Elohim (one of the main terms Jews use for our god)

soweli Lapate li lon ma weka ↩️

[Reply to:](#1162756469771472988 message) But allah is the same root as el (hebrew for God), right? Allah isn't someone's name.

[heady otter]

iirc it's a contraction of al and illah (which i think is cognate with Hebrew illoh)

checking wiktionary, illah is cognate with el and eloah

from proto Semitic *'il

[finite nest]

sewi is also just a normal toki pona word

so is God

[glass iris]

Well of course.

[lucid niche]

yeah; they're both from the Semitic triconsonantal root /ʔ-l-h/ (from PS *\ʔil or /ʔ-l/ with a vocative suffix). Allah (Arabic) was originally definite ("the god"), while Elohim (Hebrew) was originally plural ("gods"), though both are normal now. El (/ʔ-l/) was also a god of the Canaanites and ancient Mesopotamians, and is seen at the ends of names like Samuel, Daniel, Raphael, Michael, Gabriel, Ismael, and Israel.

kisa sin Pisowe pi sike sin ↩️

[Reply to:](#1162756469771472988 message) iirc it's a contraction of al and illah (which i think is cognate with Hebrew illoh)

[hollow valley]

for my linguistics class i am working on a toki pona parts-of-speech tagging process in order to analyze the distribution of different words in different syntactic positions (e.g. pona is probably more likely to be function as an adjective, pali is probably more likely to function as a verb, etc). i'm currently just thinking of ways to actually implement this

https://web.stanford.edu/~jurafsky/slp3/old_oct19/8.pdf this method looks very promising

it's an interesting problem because i want to be able to differentiate between "mi moku" and "kili li moku"

we speakers intuitively understand that "mi moku" is likely to be moku=verb since it's way less likely that i am food

and "kili li moku" is likely to be moku=noun since it's way less likely that kili is eating

but it's tricky creating something that will be able to know the difference

[timber moat]

amazing AI overview

[heady otter]

"mute"
- ilo gemini

[lucid niche]

they're all like this

jan Wekolo ↩️

[(click to see attachment)](#1162756469771472988 message)

well, likely not the last part, but

[snow echo]

"o moli e sina" - ilo Kukulu

[lucid niche]

"how many?" "many."

[rare crescent]

anywhere between 1 and technological singularity

[lucid niche]

win

[sonic leaf]

10 others guessed this too

[rare crescent]

I am surprised you didn’t know that there are new people who beat english with toki pona every week

[timber moat]

google translate accurately sharing it's feelings about toki pona

[opal hedge]

musi a

[dense nexus]

I see a lot of "laptop" in Toki pona translations on youtube

[timber moat]

some of these evoke great humor

[lucid niche]

"laptop" in Te Reo Māori is "pona"

mega b ↩️

[Reply to:](#1162756469771472988 message) I see a lot of "laptop" in Toki pona translations on youtube

(that's the indigenous language of Aotearoa / New Zealand, if anyone's wondering)

[sonic leaf]

i'll take it.. but...

[lucid niche]

win!

[runic kestrel]

that is so cool

[mystic flume]

toki pona

[lucid niche]

(it actually means "lap", the full term being "rorohiko pona". but yeah)

tokimaxxed and ponapilled ↩️

[Reply to:](#1162756469771472988 message) that is so cool

oh, and a toki is a traditional Maori axe / adze / chisel

so if you've ever heard "laptop ax", that's where it comes from

[runic kestrel]

a good rorohiko!

[meager jackal]

i was born in 1998 i don't know what a toki pona is sorry

[glass iris]

sina sona toki Mawi anu seme? mi wile kama sona ni.

[heady otter]

"I'll die today"
- my markov chain like thing

it also generated the emoji after that

[lucid niche]

mi sona ala! mi sona e ni ale tan alasa ilo

soweli Lapate li tawa mute ↩️

[Reply to:](#1162756469771472988 message) sina sona toki Mawi anu seme? mi wile kama sona ni.

[glass iris]

tenpo weka pini la, mi lon ma Nusilan la, mi esun e lipu tu pi toki Mawi. taso mi lukin ala e ona! tenpo lon la mi lon ilo waso li kama tan ma Nusilan. mi kama sin lon tomo mi la, mi wile open e kama sona pi toki ni.

[lucid niche]

o nii

[devout wedge]

https://youtube.com/shorts/BcEUXP9FUIQ?si=Kxk2lfsgOImtBxgS ai is failing cognitive tests

[dense nexus]

Im surprised about how high of a score they have

Sure not perfect, but certainly a stepping stone in the medical field if one were to create a specialized model

[devout wedge]

yes but it's still

not good

that 4o is the only one with a score that's even barely passing

[dense nexus]

Right, but gpt4o got a passing grade, which is certainly a sign of possible improvements

Obviously I would choose a human over an LLM

but still it's advancing

[devout wedge]

yes just... not much

and it's telling that the worst one is gemini

[dense nexus]

Oh Gemini always sucks

But I'm assuming the paper tested 1.5 from a couple months back

Google's newer model beats gpt4o on all benchmarks

[devout wedge]

wait really???

[dense nexus]

i think most of them

[devout wedge]

is it a gemini model or a new model

like, I remember google bard

[dense nexus]

Google bard is DOOKIE POOP bearly good for anything. But as for Google's new Gemini model, It's a new model that you can use, it's experimental so you have to go to aistudio.google.com to test it
But if you want to see it's performance, Check out https://lmarena.ai/ go to leaderboard -> overview

[devout wedge]

holy shit it says gemini 2.0 is also better than gpt 4o?

[dense nexus]

Ye

But take it with a grain of salt, usually models have different knowledge suited for different use cases

E.g. claude is really really good at programming, but has a pretty old dataset, or one model is better at multilingual tasks than the other

But anywho the point I'm getting at is, any developer can train a model to ace any benchmark, e.g. that medical benchmark, but no benchmark can fully encapsulate the complexity of actual real life use cases.

If openai wanted to make a chatgpt model suited for a hospital, they could easily do so, all they would need is a bunch of data, however, quality will always fluctuate due to the nature of LLMs

[twilit stag]

jan Misali li noka e sike

[opal hedge]

jan Jan pi ma Poka

[lucid niche]

Jan Misali is not a Polish footballer. though I'll admit I thought they were Slovak for like a year

[low wren]

Jan Miszali

Jan Miszalski

[lucid niche]

I think we misread them as Jan Mišali the first time

[low wren]

honestly same, back in the day when I thought their first name is Jan

[lucid niche]

us too, us too

[mystic flume]

toki vs pona

toki wins!

[lucid niche]

noooo

[quiet geyser]

WHAT

[meager jackal]

they are as real as their theories are

[lucid niche]

so truee

[heady otter]

real

[quiet geyser]

ah so it was a metaphor the whole time

[meager jackal]

looking back at this it's really a good way to mark tone without diacritics

[jaunty jungle]

idk man i just had to think of something and that came to mind

[meager jackal]

ma mAA maA mAaA mAa

sort of like laadan

but with capitalisation instead of á

[meager jackal]

láadan*

[jaunty jungle]

lAadan

[meager jackal]

yeah

wOoO jiAao Niwe

[lucid niche]

this is the funniest shit ever sorry

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) wOoO jiAao Niwe

dAa jia hAaO wOoO yAo chi bIing qi lIing (this is specifically how he says it)

[opal hedge]

wOoO hEeEn xi huAn bIng qi liIn

[meager jackal]

wasn't there a thing where him saying it changed the pronunciation of one of the characters because he said it wrong

wOoO? no. OwO

[opal hedge]

ō wō

[meager jackal]

wait no not the tone

it was the q in bingqiling

he pronounced it like ch

i should never doubt john xina's tone mastery

冰匙淋

honestly because qilIing was borrowed from english it could've been chiling

well no

because ki corresponds to qi

[lucid niche]

he said bing1 qi1 ling4

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) wasn't there a thing where him saying it changed the pronunciation of one of the characters because …

iirc

i dont really remember tnh

ok i dont 100% blame him for this

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) he pronounced it like ch

[meager jackal]

actually yeah he did

[lucid niche]

tones all wrong but it was iconic so

[meager jackal]

bingchiling when

[finite nest]

today i learned linguist li kon

[lucid niche]

does that n have a high tone on it? :p

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) wOoO jiAao Niwe

[meager jackal]

consonants with tones...

[lucid niche]

it was borrowed from English "cream", not "chilling". ⟨q⟩ is an allophone of /k/ before close front vowels (⟨i y⟩); "qílín" would have originally been pronounced "kílín".
this word in specific was borrowed through southern Sinitic languages, which keep the [k]; see Cantonese kei4lam4, Taiwanese Hokkien kî-lîm, and Hakka khì-lìm

just realized you said this already. anpa a

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) honestly because qilIing was borrowed from english it could've been chiling

it exists! Mandarin has marginal syllabic nasals m n ng

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) consonants with tones...

I think one of my systemmates made a list of them once, but I could not possibly tell you where it is

[meager jackal]

i know cantonese has "m" which is used very frequently

[lucid niche]

yeah, they're much more common in southern Sinitic languages

唔 is the main negator in Cantonese (m4), Hakka (m̀), Teochew (m6), and maybe Xiang (n6)